Random effects negative binomila analyses with residual diagnostic checks and robust standard errorts

[DY Kim]

Hello,

Using panel data, I conduct random-effects negative binomial analyses with the xtnbreg command.

I just wonder whether/how I can conduct residual diagnostic checks for serial correlation, heteroscedasticity, and cross-sectional dependence.

Is there any way I can estimate robust standard errors? I have found that the xtnbreg command does not allow for using the vce(robust) commend for robust standard errors.

Any advice would be greatly appreciated.

[George Ford]

Switch to xtpoisson and use vce(cluster clusvar). The robust errors abandon the poisson mean/var restriction. Then you don't have to worry about hetero or dependence.

See this thread:

https://www.stata.com/statalist/archive/2013-01/msg00272.html

[DY Kim]

Switch to xtpoisson and use vce(cluster clusvar). The robust errors abandon the poisson mean/var restriction. Then you don't have to worry about hetero or dependence.

See this thread:

https://www.stata.com/statalist/archive/2013-01/msg00272.html

George,

Thank you for the advice. I used xtpoisson with robust standard errors.
Number of obs = 107,763
Number of groups (ct) = 2,113
Observations per group = 51 time series
Ct is nested within 5 dummy area variables

I ran the random effects poisson model with robust standard errors clustered on ct (group variable) as follow: xtpoisson count dummy1 dummy2 timeconstant1 timeconstant2 b1.area, re vce(cluster ct). After running the above model, I found that all dummy and timeconstant variables are statistically significant.

I ran the xtpoisson model by changing a cluster variable from ct to area: xtpoisson count dummy1 dummy2 timeconstant1 timeconstant2 b1.area, re vce(cluster area). Ct is nested within 5 dummy area variables. Once I changed clusvar from ct to area, all timeconstant variables turned out to be insignificant, while two dummy variables are still significant. Would you help me to interpret this? Why did the timeconstant variables become insignificant?

For comparison reasons, I ran the pooled poisson model using area as a cluster variable: poisson count dummy1 dummy2 timeconstant1 timeconstant2 b1.area, re vce(cluster area). All dummy and timeconstant variables are significant.

Why do all timeconstant variables become insignificant only in the random effects poisson model with robust standard errors clustered on 5 areas? Can you help me understand this?

The number of areas is only 5. There is a concern over having too few clusters. Can I still use area as a cluster variable? Which one should I use between ct and area for robust clustering errors?

Any advice would be greatly appreciated.

[Jeff Wooldridge]

You should not be clustering with five clusters. The statistical properties are based on the number of clusters increasing. Clustering with five clusters is like saying you only have five independent pieces of information. You can cluster at the ct level.

I would avoid the random effects poisson estimator as it is not robust to any violations of assumptions. Fixed effects poisson is fully robust. Is there a reason you're interested in the time-constant variables?

[DY Kim]

Is there a reason you're interested in the time-constant variables?

Hello Jeff,

Thank you very much for the information. It helps me understand better statistical implications of using few clusters. The reasons I wanted to use random effects modeling is to examine how much time constant variables are related to changes in the dependent variable, even if they could not be said to cause any changes in the dependent variable over time. Fixed effects models can statistically adjust for all time-invariant variables between the groups but can't give us any information how they associated with changes in the dependent variable. In my discipline, many people are interested in knowing such information. Any advice would be appreciated. Thank you for sharing your expertise.

[DY Kim]

I would avoid the random effects poisson estimator as it is not robust to any violations of assumptions. Fixed effects poisson is fully robust. Is there a reason you're interested in the time-constant variables?

I ran the fixed effects poisson models. I found that the number of observations dramatically decreased from 107,763 in the random effects poisson model to 55,080 in the fixed effects poisson models. Is this because all time invariant variables were removed from the models? Would you please provide some insight on this? Thank you.